Don't Always Pick the Highest-Performing Model: An Information Theoretic View of LLM Ensemble Selection